Claudeforce at Dreamforce 2026: What Salesforce Concedes by Naming a Model Maker

Salesforce now ships a product with another company's model in its name. Claudeforce put its first piece, a plugin carrying 37 sales skills, into open beta at Dreamforce 2026, and Salesforce has published no price for it. The beta is the least interesting part. The interesting part is what a CRM vendor concedes when the brand on the box belongs to Anthropic.

The short answer

Claudeforce is a Salesforce and Anthropic partnership. Its first product, Salesforce in Claude, is a plugin that lets Claude read and act on Salesforce records under each user's existing permissions. Salesforce says it is available to all customers in beta from 15 September 2026. No price, service level or general availability date has been published.

This post is written for the head of revenue operations deciding whether to put the beta in front of sellers. That reader owns the rollout, the permissions and the renewal, so the questions below are the ones that land on their desk.

What Salesforce actually shipped at Dreamforce 2026

Dreamforce 2026 ran from 15 to 17 September in San Francisco. The keynote launched AIforce, which Salesforce describes as "a live interface layer that brings the full power of Salesforce to wherever people and agents work". It launched with three surfaces: Claudeforce, Slackforce and Agentforce Coworker.

Strip away the branding and three separate things happened. Each sits at a different stage, and treating them as one announcement is the most common mistake in the recap coverage.

Salesforce in Claude, the plugin

This is the product with the model's name attached. Salesforce and Anthropic announced it on 26 August 2026 as a plugin with 37 prebuilt sales skills, available then to select pilot customers. On 15 September the AIforce story said it had been piloted by Deloitte, GitLab and Legora and is now available to all customers in beta.

The skills run inside Claude and reach Salesforce through Headless 360, which the Claudeforce page says AI can call directly over MCP. Patrick Stokes, Salesforce's president of applications, told CIO on 15 September that customers can grab the plugin in open beta and install it into a Claude enterprise account.

Claude inside Agentforce

This runs the other way: Salesforce's agents calling Anthropic's model. The August release says Claude serves as a reasoning model for the Atlas Reasoning Engine, powers Agentforce Vibes and Agentforce Coworker by default, and is available in Agent Builder. It also repeats a claim first made in October 2025, that Claude is the first LLM provider fully integrated within the Salesforce trust boundary, reached through Amazon Bedrock.

This part is not new and it is not beta. Anthropic's own October 2025 post said Agentforce powered by Anthropic was available that day for select customers, with CrowdStrike and RBC Wealth Management named as users.

Koa, Salesforce's own reasoning model

The third launch points in the opposite direction. Koa is a CRM reasoning model built by post-training NVIDIA's Nemotron 3 Super on synthetic Salesforce workflow data. The 15 September press release says Salesforce controls the weights and runs post-training and inference inside its own trust boundary.

Koa is with select pilot customers now, with general availability expected in winter 2026 in US regions. Its headline claim, three times fewer errors on CRM actions, comes from Salesforce's own benchmark. Tech Times reported on 17 September that the methodology has not been published, so treat the figure as a vendor claim until it is.

The Claudeforce ledger: shipped, beta, pilot or unpriced

Read the releases with one question per line item: can a customer buy this today, at a known price, with a support commitment? Almost nothing announced at Dreamforce 2026 clears all three.

None of this is unusual for a conference. Vendors announce ahead of availability every year. It matters here because the beta asks for something real: an admin connection that lets a third-party model read and write your system of record.

Why naming Claude inside the product is a concession

Salesforce has spent two years telling customers its agents are model-agnostic. Agentforce routes reasoning through Atlas, and the model underneath is a configuration choice. Claudeforce breaks that pattern. The model vendor's name is now the product name, and the product lives in the model vendor's app.

Our read is that this is a deliberate concession, and a sensible one. It is also a concession that would have been unthinkable in Salesforce's marketing two years ago.

The brand on the box tells you who owns the session

In Salesforce in Claude, the seller opens Claude, not Salesforce. The daily habit, the chat history and the place where work starts all belong to Anthropic's product. Salesforce becomes something Claude calls.

Benioff's own line for this is "Here, the UI is the AI". Stokes put it more sharply at Dreamforce, saying other software companies think their product is the UI, while Salesforce's product is the trust customers place in it to hold their data. That is a company redefining its product so that losing the screen does not count as losing.

We think the redefinition is honest about the direction of travel. It also gives up the most defensible position a SaaS vendor has historically held, which is being the place a user spends the day. That shift is the subject of our earlier analysis of what happens to software when the dashboard stops being the product.

Default model is not exclusive model

The release language is precise, and worth reading slowly. Salesforce is Anthropic's "preferred CRM". Claude is the "default model" for Slack AI, Slackbot, Salesforce in Claude, Headless 360, Agentforce Coworker and Claude Code across Salesforce engineering. Nowhere does the release say "exclusive".

Defaults can be changed. PPC Land noted on 30 August that neither company disclosed contract value, duration, minimum commitments or exclusivity terms. The commercial relationship runs both ways too: the release says the two companies are strategic customers of one another. The Next Web reported that Salesforce expects to spend around $300 million on Anthropic tokens this year and holds a stake it values near $5 billion. Those two figures are TNW's, not Salesforce's, and we could not trace them to a filing.

Where agent value sits: the CRM case Salesforce is making

Every Dreamforce 2026 message about AIforce makes one argument. Models reason, but they do not know what is allowed. The value, Salesforce says, is in the layer that does.

The September takeaways piece states it plainly: models "can show what's probable, but they can't determine what's allowed". Agents, it continues, need the APIs, metadata, business semantics, permissions and governance Salesforce has spent 27 years building.

Permissions are the strongest part of the argument

The Claudeforce page says Claude sees what the user is authorized to see, and can do what that user is authorized to do, no more. An admin connects once and the plugin inherits the existing permission model. There is no per-user setup and no new permission model to audit.

This is a genuine advantage, and it is the part of the pitch we find most convincing. Rebuilding record-level access rules in a new agent platform is slow and error-prone. Our piece on agent identity and non-human access covers why that rebuild is where many agent projects stall. A vendor that removes it removes a real cost.

The Trust Layer was built for the other direction

Salesforce's trusted AI page describes a Trust Layer with zero data retention, toxicity detection, and data masking that replaces sensitive data with tokens before the prompt is sent to the LLM. That design assumes Salesforce assembles the prompt and calls the model.

In Salesforce in Claude, the prompt starts in Claude and Claude calls Salesforce. The Claudeforce page promises zero data retention on Sonnet, Opus and Haiku, and says Anthropic's Enterprise Frontier Safeguards roll out in phases beginning later this fall. Neither page we read says whether Trust Layer masking applies to records returned to Claude over MCP. That is not evidence of a gap. It is a question the documentation does not yet answer, and you should ask it in writing.

Koa and the model-side hedge

If the value really sat only in data and permissions, Salesforce would not need its own model. It built one anyway, and launched it the same day as AIforce.

The Koa release quotes Benioff saying that "the most valuable thing Salesforce has built isn't our platform", but the accumulated knowledge of how enterprise business works. He adds that with Koa, that knowledge is put inside the model itself. On the same stage, Stokes was arguing the value lives outside the model, in data and metadata.

Both cannot be the whole story. We read Koa as an insurance policy. If frontier models commoditise, Salesforce wants a cheap model it controls for routine CRM actions. If frontier models keep pulling ahead, it wants Claude as the default. Running both lets it avoid choosing. Our earlier post on the real switching costs between model providers explains why that optionality is worth paying for.

The OpenAI thread makes the same point. Salesforce's October 2025 release with OpenAI said OpenAI would be featured as a preferred model in Agentforce, selectable as the default for Atlas and Prompt Builder. On 16 September 2026, Salesforce said Missionforce plans to integrate OpenAI's frontier models through Amazon Bedrock and make its workflows available through ChatGPT. A consultancy preview reported a Salesforce email saying Sam Altman would join Benioff on the main stage. We found no newsroom transcript of what he said, so we do not characterise it.

How Salesforce got from preferred model to plugin

The naming decision looks abrupt only if you read the August release on its own. Laid out as a sequence, it is the third step in a 12-month move from model as component to model as storefront.

The October 2025 Anthropic post said Claude would be a preferred model for Agentforce in financial services, healthcare, cybersecurity and life sciences, with all of Claude's traffic contained within the Salesforce virtual private cloud. The same day, Salesforce published its OpenAI release using the same word. In April 2026, PPC Land reports, Agentforce Vibes 2.0 supported both the Anthropic and OpenAI agent SDKs.

So the August move is not Salesforce picking a winner. It is Salesforce letting one partner put its own name on a storefront, while every other door stays open. That is a distribution deal more than a technology choice. It sits closer to the dynamics in our analysis of who captures value in agent marketplaces than to a platform migration.

Our position: the CRM reading is right about this year and the model reading is right about the direction. The price, when Salesforce publishes it, will tell you which one Salesforce itself believes.

Where this reading is weakest

Three parts of this argument rest on less than we would like. Here they are.

We are reading a marketing record, not contracts

Everything above comes from press releases, a product page, newsroom stories and trade coverage. We tried to read Salesforce's beta services agreement, linked from its customer agreements page, and could not extract the text from the published PDF. So we do not quote its terms. The release footer says only that availability is governed by customer agreements and that purchasing decisions should rest on products currently available.

The financial figures that would settle the value question are not public either. Token spend and stake value come from one outlet. Salesforce does not break out Agentforce model costs by provider.

The concession may be smaller than it looks

A fair counter-case: Salesforce has run integrations into other people's interfaces for years, including Slack, which it owns, and Outlook and Gmail, which it does not. A Claude plugin may be one more integration, not a strategic surrender. On that view, naming the product after Claude is co-marketing, and the CRM keeps the margin as it always has.

We think that view underrates the difference. An email plugin surfaces records. This plugin reasons over them, writes to them and chooses which skill to run, inside an app Salesforce does not control. But the counter-case could be right, and the first pricing announcement is the evidence that would show it. The prior case that Salesforce survives the agent shift intact is argued in full in our analysis of whether Salesforce is a platform or roadkill in the agent era.

What to check before you join the Claudeforce beta

A beta that connects a third-party model to your system of record deserves the same questions as a production integration. The absence of a price does not mean the absence of cost: admin time, permission review and seller habit are all spent on day one.

Two of those rows connect to work we have done elsewhere. For the contract side, our comparison of model provider terms on retention and training covers what Anthropic's own enterprise terms already commit to. For the cost side, our breakdown of Agentforce conversation pricing shows how the existing meter behaves, which is the best guide to how a future Claudeforce price might be structured.

If we ran revenue operations at a mid-sized Salesforce customer, we would enable the beta for one team, with read-only skills first, and refuse write access until the logging question has a written answer.

Frequently asked questions

How to run a two-week Claudeforce test

Pick one sales team of five to ten sellers who already use Claude. Enable only read skills such as pipeline review and call prep, and log one number before you start: how long weekly pipeline review takes today. Compare it at the end of week two.

In the same fortnight, send your account team the five questions in the table above, in writing. Whatever comes back, especially on masking and logging, is worth more than the test itself. A vendor that can answer all five has shipped a product. A vendor that cannot has shipped a beta, and you should treat it as one.

Related reading